A context window is the maximum amount of text a language model can read at once, measured in tokens. A token budget is the share of that window a system chooses to spend on retrieved web content when answering a question. Even when context windows are very large, AI search systems pass the model a small selection of passages, so only the most relevant part of your page is likely to be read.
What a token is
Language models split text into tokens: whole words, parts of words or punctuation. OpenAI’s guidance is that one token is roughly four characters of English text, or about three quarters of a word. The word “optimisation” may be two or three tokens; “SEO” may be one.
Token counts depend on the language. Tokenizers are usually trained mostly on English text, so languages such as Hindi and Marathi often need more tokens for the same meaning. For sites that publish in Indian languages, the same answer takes up more of the model’s budget.
Context window versus token budget
Model context windows have grown from a few thousand tokens in early models to hundreds of thousands, and in some models more than a million. That does not mean an assistant reads whole websites when it answers. Every token costs computing time and money and adds delay, and longer inputs raise the risk of the model missing the relevant part, as the lost-in-the-middle research showed.
So search systems set a budget. They retrieve many candidates, re-rank them, and pass only the top passages to the model, often a few hundred words from each source. The exact budgets are not published and vary by product and question.
What gets cut
When a crawler or retrieval system processes a page, most systems strip navigation, footers, cookie notices and scripts, then split the remaining text into chunks. From those chunks only the best matches are kept. Text that is likely to be dropped includes:
- Long introductions that do not answer the question
- Boilerplate repeated on every page, such as company descriptions in every footer
- Text inside images, which a text crawler cannot read
- Content loaded only after clicks or scrolling
How to write for a limited budget
State the answer early in each section
If only one passage from your page is used, it should be the one that answers the question, and it should do so in its first sentence.
Cut filler
Phrases like “in this article we will explore” use tokens and add nothing a model can use.
Keep each section self-contained
A passage that begins “As mentioned above” is less useful when the text above it was not selected. Name the subject again.
Put numbers and conditions next to the claim
A sentence such as “The basic hosting plan costs 800 rupees a month for sites under 10 GB” survives being cut into a small passage. A price in one paragraph and its conditions three paragraphs later may not.
Frequently asked questions
Do bigger context windows mean long pages will be read in full?
Not in AI search. Systems still select passages to keep answers fast and affordable. Long pages help only when each section is useful on its own.
How many tokens does my page use?
Tokenizer tools from OpenAI and others will count tokens for any text. As a rough guide, a 1,500-word English page is around 2,000 tokens before navigation and other page text is removed.